Incompatible with transformers 5.x: find_pruneable_heads_and_indices removed

[vittorio-prodomo]

Description

Surya 0.17.1 fails when installed with transformers>=5.0. Two separate breakages:

1. Donut encoder import error

The table recognition module imports find_pruneable_heads_and_indices from transformers.pytorch_utils, which was removed in transformers 5.x:

surya/surya/common/donut/encoder.py

Lines 8 to 12 in e735028

	from torch import nn
	from transformers.activations import ACT2FN
	from transformers.pytorch_utils import (
	find_pruneable_heads_and_indices,

ImportError: cannot import name 'find_pruneable_heads_and_indices' from 'transformers.pytorch_utils'

2. Layout predictor config attribute error

The layout predictor accesses bbox_size on the model config, which no longer exists under transformers 5.x:

surya/surya/layout/__init__.py

Line 16 in e735028

batch_size = settings.LAYOUT_BATCH_SIZE

AttributeError: 'EfficientViTConfig' object has no attribute 'bbox_size'

Steps to reproduce

pip install surya-ocr==0.17.1 transformers==5.3.0
python -c "from surya.table_rec import TableRecPredictor"

Root cause

pyproject.toml specifies transformers >= "4.56.1" with no upper bound, so pip/uv will resolve to 5.x:

surya/pyproject.toml

Line 28 in e735028

pre-commit = "^4.2.0"

Workaround

Pin transformers>=4.56.1,<5 in your project dependencies.

Environment

• surya-ocr 0.17.1
• transformers 5.3.0
• Python 3.12.12
• torch 2.10.0
• Linux x86_64

Suggested fix

Either add an upper bound (transformers>=4.56.1,<5) to pyproject.toml, or update the imports/config access to use the transformers 5.x API.